At Rise of AI 2019 in Berlin, Ulrich Schmidt, Segment Manager at EBV Elektronik, a semiconductor distributor, explains how his company helps customers turn AI ideas into products running on real hardware, from microcontrollers up to FPGA-based accelerators. He describes EBV as a roughly two-billion-dollar distributor of active semiconductor components with about 1,000 employees and 370 million parts in inventory, serving mainly mid-sized industrial, automotive and smart-consumer customers across 59 offices in Europe, Africa and Russia. He surveys where customers apply AI, including retail shelf monitoring, manufacturing predictive maintenance, medical image diagnosis, financial anomaly detection and assisted driving, then walks through AI-relevant product lines from three key suppliers: NXP's edge microcontrollers and crossover processors, Xilinx's FPGA accelerator cards for data centers, and STMicroelectronics' toolchain for converting trained models onto microcontrollers. He frames EBV's own role as focused on edge inference of pre-trained models rather than the cloud-based training side of AI, closing with a short Q&A on why dedicated AI chips remain limited among the distributor's supplier lineup.
Key takeaways
- Schmidt positions EBV Elektronik as focused on helping customers deploy pre-trained AI models at the edge, not on the cloud-based training side of AI.
- He gives predictive maintenance, such as detecting a failing ball bearing from sensor data before an unplanned shutdown, as a concrete manufacturing use case.
- NXP's low-cost "crossover" processors, priced around three to four dollars, are highlighted as well suited to running edge AI algorithms.
- Xilinx FPGA cards such as the U200, U250 and U280 target data-center AI acceleration and are already used in AWS's F1 cloud instances.
- STMicroelectronics' X-Cube-AI tool converts trained models from frameworks like TensorFlow, Caffe or Lasagne to run directly on its microcontrollers.
Detailed notes
EBV Elektronik and its markets
Schmidt introduces EBV Elektronik as a distributor of active semiconductor components, part of the Avnet group, which also includes Silica and divisions selling passive and electromechanical components and server-rack parts. He describes EBV as handling the full customer journey from design and engineering recommendations to sourcing and logistics, comparing its logistics role to Amazon for active components. He gives EBV's scale as roughly a two-billion-dollar company with about 1,000 employees, just over 100 technical engineers, about 200 field sales staff, and an inventory of about 370 million parts held in a warehouse near Munich, with headquarters also in the Munich area. Its main suppliers include NXP, ST Micro, Microchip and Renesas, and it sells across 59 offices in Europe, Africa and Russia, primarily to mid-sized industrial customers, followed by automotive, smart consumer, smart grid, and high-reliability healthcare and medical segments; larger customers, he notes, are typically supplied directly by the chip makers.
Where customers apply AI
Schmidt surveys application areas he sees customers exploring: in retail, monitoring whether shelves are empty and using sentiment or demographic analysis, such as gender and age, to customize digital displays and advertising; in manufacturing, predictive maintenance that uses sensor data to detect, for example, a ball bearing likely to fail within weeks so maintenance can be scheduled before an unplanned shutdown; in medical imaging, helping flag cancer risk from diagnostic images to advise a doctor; in finance, detecting anomalies in fast trading; and in automotive, assisted driving, where he describes modern cars as effectively data centers generating large data streams that need processing.
NXP: edge microcontrollers and crossover processors
Schmidt describes NXP's range from Cortex-M4-based microcontrollers up to multi-core devices with as many as 16 A53 cores aimed at gateway and network applications, plus GPU-equipped devices for graphics tasks such as decoding video up to 4K. He highlights NXP's "crossover" devices, a hybrid between microcontroller and microprocessor priced around three to four dollars and running from 600 megahertz up to about one gigahertz, which he says are well suited to running AI algorithms; NXP is also adding dedicated IP blocks to accelerate AI functions and provides a software ecosystem for deploying pre-trained models. He notes EBV's presented solutions focus on edge processing rather than training, converting and running pre-trained models on-device for post-processing of sensor data.
Xilinx: FPGA acceleration for the data center
Schmidt presents Xilinx, comparable in size to Intel's programmable solutions business, as targeting data-center applications such as fintech fast trading, database acceleration, video transcoding and machine learning, with FPGAs allowing algorithms to be tailored and quickly adapted as new, more efficient neural network designs emerge. He describes the U200 and U250 PCI Express accelerator cards, the U280 with added high-bandwidth memory for memory-intensive workloads, upcoming network-focused cards, and a forthcoming 7-nanometer chip called Versal, combining universal and versatile building blocks optimized for AI. Xilinx's ecosystem also offers model pruning and guidance on partitioning workloads between standard x86 processing and the FPGA; he notes AWS's F1 cloud instances already use Xilinx FPGA cards to accelerate AI workloads.
STMicroelectronics: converting models for microcontrollers
Schmidt describes STMicroelectronics' Cube ecosystem and its X-Cube-AI extension, which takes trained models from frameworks such as TensorFlow, Caffe or Lasagne, converts them, and integrates them into ST's development environment to run on its microcontrollers, spanning entry-level Cortex-M0 devices up to Cortex-M7 devices running roughly 300 to 400 megahertz with extensive peripheral support.
Q&A highlights
Asked why dedicated, application-specific neural chips seem hard to find outside obscure Chinese suppliers, Schmidt says the answer partly reflects EBV's own supplier lineup, and that he is not fully familiar with competitors' offerings; he notes Nvidia is typically used for training workloads that need floating-point operations, while inference workloads can often use smaller bit sizes, and says that for most of EBV's suppliers, AI capability is currently built into their existing product lines rather than offered as dedicated standalone chips.
Questions answered in this talk
What kind of AI hardware does EBV Elektronik focus on?
According to Ulrich Schmidt at Rise of AI 2019, EBV Elektronik focuses on edge processing of pre-trained AI models, from low-cost microcontrollers up to FPGA-based accelerator cards, rather than the cloud-based training side of AI.
What is an example of predictive maintenance using AI mentioned at Rise of AI 2019?
Ulrich Schmidt described using sensor data to detect, for example, a ball bearing that is likely to fail within weeks, allowing maintenance to be scheduled ahead of time rather than reacting to an unplanned production stoppage.
How does STMicroelectronics help developers deploy trained AI models on microcontrollers?
At Rise of AI 2019, Ulrich Schmidt explained that ST's X-Cube-AI tool, part of its Cube ecosystem, converts trained models from frameworks such as TensorFlow, Caffe or Lasagne and integrates them into ST's development environment to run on its Cortex-M microcontrollers.